{ "intent": "log_meal", "meal_type": "lunch", "items": [ {"food_name": "Roti", "grams": 96}, {"food_name": "Dal", "grams": 150}, {"food_name": "Sabzi", "grams": 130} ] }

It is optimized to:

Output ONLY valid JSON
Avoid commentary
Normalize quantities
Multiply counts × grams automatically

Downstream Use

Suitable for:

Nutrition apps
Macro tracking systems
Health dashboards
AI nutrition assistants
API backends requiring structured output

Out-of-Scope Use

Not intended for:

Medical diagnosis
Calorie accuracy guarantees
Free-form conversational AI
Open-ended creative writing
Sensitive decision making

Training Details

Training Data

~20,000 structured nutrition logging examples
Instruction-style prompts
JSON-formatted targets
Explicit "JSON-only" reinforcement

Data focused on:

Meal type detection
Food item normalization
Gram calculations
Structured intent extraction

Training Procedure

QLoRA 4-bit fine-tuning
LoRA adapters applied to attention layers
Mixed precision training
Adapter merged into base model post-training

Training Regime

4-bit quantization during training
Full precision merged model for inference
Supervised fine-tuning (SFT)

Inference Example

python

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "Shreyasjani/nutribrain-mistral7b-merged"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    device_map="auto"
)

prompt = "Return ONLY JSON.\n\nLog dinner: 150g paneer, 110g cooked brown rice"

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

output = model.generate(
    **inputs,
    max_new_tokens=200,
    temperature=0.1,
    do_sample=False
)

print(tokenizer.decode(output[0], skip_special_tokens=True))

Recommended settings:

temperature: 0.1 – 0.3
do_sample: False
top_p: 0.9
max_new_tokens: 150–300

Bias & Limitations

Nutrition values are not verified against real databases.
Assumes grams when unspecified.
English-only.
Sensitive to prompt formatting.
Requires explicit "Return ONLY JSON" instruction for best results.

Technical Specifications

Architecture: Mistral 7B
Parameters: ~7 billion
Context Length: 8k
Format: Safetensors
Compatible with: Transformers, TGI, vLLM

Deployment

Recommended:

Hugging Face Inference Endpoint (TGI)
vLLM server
Dedicated GPU inference

Minimum GPU:

A100 40GB (recommended)
H100 (ideal)
24GB GPU may work with optimized inference

Environmental Impact

Hardware: NVIDIA H100 / A100
Fine-tuning duration: ~20k steps
Adapter size: 168MB
Merged model size: ~14.5GB

Author

Shreyas Jani
Fintech & AI Systems Builder

Contact

Hugging Face: https://huggingface.co/Shreyasjani